Skip to content
Open Snotel home
MCP

Examples

Worked agent tasks over MCP: latest snow depth near a point, a week at three stations, the largest gains in snow water equivalent, how unusual today is, a water year against history, the SNOTEL stations for a ski resort, and freshness.

Each example is a question, the tool calls an agent makes to answer it, and how to read what comes back. The tools return structured data from the stored observations; anything computed from them, such as a difference between two readings, is the agent's own arithmetic and should be described as that. The tools are listed on the tools page.

Find stations near a place and show their latest snow depth

"Find stations near this coordinate and show their latest snow depth."

The agent starts with the stations nearest the point:

{
  "tool": "stations_nearest",
  "arguments": { "lat": 44.31, "lng": -115.2, "limit": 3 }
}

Each result has a station id (its triplet) and distanceKm. It then asks each one for snow depth, SNWD:

{
  "tool": "station_latest",
  "arguments": { "station": "312:ID:SNTL", "elements": ["SNWD"] }
}

Read each reading's freshness before saying it is current. A fresh zero means there is no snow at that station; a stale reading means no recent data has reached Snotel, so the value is old whatever it says. observedAt is when the reading's period began, and a unit accompanies every value.

Compare the last seven days at three stations

"Compare the last seven days at these three Idaho stations."

If the agent does not know the stations' ids, it searches by name and state:

{
  "tool": "stations_search",
  "arguments": { "query": "banner", "state": "ID", "active": true }
}

It can check which measurements a station reports before asking for them:

{ "tool": "station_get", "arguments": { "station": "312:ID:SNTL" } }

Then it reads the week from each station. Dates are the station's local calendar days, inclusive:

{
  "tool": "station_observations",
  "arguments": {
    "station": "312:ID:SNTL",
    "elements": ["WTEQ", "SNWD"],
    "from": "2026-10-17",
    "to": "2026-10-23"
  }
}

The result groups readings by sensor. A reading with value: null is a day the source reported nothing for, and qualityFlag says whether it was marked suspect; the agent should report gaps as gaps, not fill them in or treat them as zero. When originalValue is present, the source edited or blanked that reading after first receiving it.

Which Idaho stations gained the most snow water equivalent in 24 hours

"Which Idaho SNOTEL stations gained the most snow water equivalent in the last 24 hours?"

snow_rankings ranks them from changes Open SNOTEL has already computed:

{
  "tool": "snow_rankings",
  "arguments": { "kind": "wteq_gain_24h", "state": "ID", "limit": 5 }
}

Each entry has the station, the gain with its unit, and an explanation naming the two readings it was computed from. Only stations whose readings are fresh rank, so a short or empty list means few stations have recent data, not that none gained. Say the result as "snow water equivalent rose by this much", not "this much snow fell": the change also reflects settling, melt and sensor behavior, and Snotel does not estimate snowfall. For snow depth, snwd_gain_24h ranks the same way, and its gain is not snowfall either. See provenance.

What the weather is doing at a station

"What's the temperature at Banner Summit, and what's the forecast?"

{ "tool": "station_weather", "arguments": { "station": "312:ID:SNTL" } }

Start from observed.primary: its kind says whether the temperature was measured at the SNOTEL station (snotel_observation) or at a nearby NWS station (nws_observation, with that station's distance and elevation), and its freshness how recent it is. Then forecast.periods, which are nws_forecast: say "the National Weather Service forecasts", never "it is". Mention any alerts.items by their headline, as NWS issued them. If forecast is null or a part is stale, say so rather than filling in.

How unusual is today's snowpack at a station

"How unusual is today's snow water equivalent at Banner Summit?"

{ "tool": "station_context", "arguments": { "station": "312:ID:SNTL" } }

In the WTEQ element, history.percentOfMedian is today's value as a percent of the station's median for this day over WY1991–2020, history.percentile where it falls among those years, and history.rank its place (of years plus today). Report them as what they are: "118% of the station's 1991–2020 median for this date, the 72nd percentile", not "118% of normal". When a field is unavailable, give its reason in words: zero_median early in the season, insufficient_samples at a station with a short record. When it is stale, say how old it is.

Chart a water year against its history

"Give me the current water year series and the historical band for Banner Summit."

{
  "tool": "station_climatology",
  "arguments": { "station": "312:ID:SNTL", "element": "WTEQ", "previous": true }
}

days has every day from October 1 to September 30 with the observed value and the day's band (p10, p50, p90). Days with no reading, including the days still ahead, have value: null: draw them as gaps. Today's entry is the newest hourly reading from today when it has observedAt; every other value is a daily (local midnight) reading. previous holds last year's values for comparison. About 365 days come back; summarize the season rather than listing them.

Which SNOTEL stations tell me about a ski resort

"How is the snowpack at Bogus Basin?"

The agent finds the resort, then asks which stations relate to it:

{ "tool": "resorts_search", "arguments": { "query": "bogus", "country": "US" } }
{
  "tool": "resort_stations",
  "arguments": { "resort": "bogus-basin-mountain-recreation-area" }
}

A curated entry was reviewed, and its relationship and rationale say how the station relates; a representative one can stand for the resort. A candidate is only one of the nearest stations, ranked by distance, and the answer should say "the nearest SNOTEL station, 1.4 km away and 174 m higher", not "the resort's station". The agent passes a triplet to station_context for its snowpack, and credits the resort data with the result's license.attribution. For the resort's own details (elevation, website, status), resort_get:

{
  "tool": "resort_get",
  "arguments": { "resort": "bogus-basin-mountain-recreation-area" }
}

Check how fresh and how well checked a value is

"Is the snow depth at Banner Summit current, and has anyone reviewed it?"

{
  "tool": "station_latest",
  "arguments": {
    "station": "312:ID:SNTL",
    "elements": ["SNWD"],
    "duration": "HOURLY"
  }
}

freshness.state answers the first half. For the second, qualityFlag and flag are the source's letters, unchanged, and isProvisional is true only when the source marked the reading provisional and false only when it marked it reviewed. Quality explains the letters.

Decide whether to trust a station's readings right now

"Can I trust what Banner Summit is reporting today?"

{ "tool": "station_data_quality", "arguments": { "station": "312:ID:SNTL" } }

For snow water equivalent, snow depth, precipitation and air temperature, the result gives the newest reading of each sensor with its freshness, how many readings in the last 7 days (daily) or 48 hours (hourly) failed a plausibility check and which checks, any known gap in knownMissing, and reasons: short sentences naming what is off. There is no score. An empty reasons means every sensor is fresh and nothing recent was flagged; otherwise say what the reasons say, such as "the snow depth has not updated in two days" or "one hourly snow water equivalent reading yesterday looked like a sensor spike", rather than quoting the value as current.

Ask for a long history

"I want every daily reading since 2000 for Idaho."

That is more than one call can return, and MCP is deliberately bounded. The agent should say so, then read one station at a time in consecutive ranges of at most 366 days of daily data, for example:

{
  "tool": "station_observations",
  "arguments": {
    "station": "312:ID:SNTL",
    "from": "2000-01-01",
    "to": "2000-12-31",
    "elements": ["WTEQ"]
  }
}

Then it repeats with the next year. The REST API takes the same ranges.